How Is Google DeepMind’s New Robotics AI Model Different From Gemini?

Compare action output, hardware range and partner access to plan a realistic robot pilot now.

Google DeepMind's new robotics AI model differs from Gemini because it outputs physical actions, not just digital content. A vision-language-action model turns camera views and language instructions into movement, so it drives motors directly while Gemini produces text, code and media, according to Google DeepMind in the launch announcement.

The current version is Gemini Robotics 2.5, introduced Nov. 10, 2026. It targets factory and automation teams that need robots to learn varied handling skills quickly.

Table of Contents

From digital answers to physical actions

Gemini answers questions and creates digital material. Robotics 2.5 converts camera frames plus natural-language instructions directly into motor control.

That shift lets one model link what it sees to how it should move. For readers, the practical difference is clear: Automation teams should therefore judge Gemini on answer quality and judge Robotics 2.5 on task completion.

  • Gemini: receives prompts and returns text, code or media.
  • Robotics 2.5: receives views and instructions and returns joint and gripper motion.

Why are there two robotics models?

DeepMind released two companion models with different jobs. Gemini Robotics controls the robot itself. Gemini Robotics-ER handles embodied reasoning instead.

Embodied reasoning means spatial understanding, pointing and task planning. Robotics-ER connects to a robot's existing controller rather than sending motor commands itself. Teams with a reliable controller can add reasoning without replacing motion hardware.

What is new in Robotics 2.5?

Robotics 2.5 increases both the breadth of learnable skills and the speed of acquiring them. DeepMind evaluated it across more than 200 distinct tasks covering delicate objects, tools, flexible materials and precision assembly. The hardware range is wide.

It is described as able to operate tabletop arms through full humanoids and 22-degree-of-freedom dexterous hands, according to Google DeepMind on the model page. Dexterous hands mean multi-fingered grippers for fine manipulation. That breadth matters for mixed workcells. A team can test picking, tool use and assembly with the same model family instead of training separate controllers.

How does it perform on active lines?

In an Apptronik-Mercedes-Benz proof of concept, Apollo 2 humanoids running Robotics 2.5 sorted metal brackets from bulk bins into sequenced kitting containers in Hungary and Germany at a reported 96% success rate, according to Unite.AI in the proof-of-concept report. Bulk bins mean unsorted parts piled together. Sequenced kitting means placing the right parts in order for assembly. Physical safety uses inference-time reasoning guidance to steer around hazards.

DeepMind reported safe task completion up 33% around obstacles and 43% around nearby people in evaluations. Those gains matter most where people and robots share aisles. Access remains limited. Robotics 2.5 is rolling out only to robotics partners such as Agile, Apptronik and Boston Dynamics, according to Google DeepMind in the version 2.5 announcement. Factories and developers must therefore contact a partner for a pilot rather than using a public Gemini-style app or API.


You Might Also Like